OpenAI models coming to Amazon Bedrock: Interview with OpenAI and AWS CEOs

Original Article ↗ Hacker News Discussion ↗

Product & Technical Details

  • OpenAI’s latest models, including Codex and “Managed Agents powered by OpenAI,” are coming to Amazon Bedrock in limited preview.
  • Bedrock’s Mantle inference engine is highlighted as being built by a small senior team and now exposes OpenAI‑compatible endpoints, easing migration by just changing base URL and key.
  • Customers still do not get raw model weights; this is managed inference, potentially “air‑gapped” inside their AWS account.

Enterprise, Privacy & Compliance

  • Many commenters say this is primarily for organizations that don’t want a direct relationship with OpenAI but already trust AWS and have DPAs and procurement processes in place.
  • Running models inside the customer’s AWS account, under AWS contracts, is seen as a major compliance win (fewer subprocessors, clearer data boundaries, HIPAA/GDPR concerns).
  • AWS states Bedrock inputs/outputs are not shared with model providers or used to train base models.
  • There is confusion over OpenAI’s legal obligation to retain logs; one comment asserting indefinite retention is directly corrected by another citing updated OpenAI language restoring 30‑day deletion.

AWS vs Azure & Go-to-Market Strategy

  • Several enterprise users report poor experiences with Azure OpenAI: slow rollout of frontier models, tight quotas, high latency/outages, and bureaucracy in getting capacity.
  • Others, especially in some regions, report Azure + OpenAI is working fine and widely adopted.
  • Multiple comments argue Anthropic’s Bedrock presence gave it a big enterprise edge; OpenAI is “catching up” by dropping Azure exclusivity and changing its Microsoft deal.

Model Quality, Benchmarks & Serving Differences

  • Commenters stress that the “same” model can behave differently across platforms due to quantization, batching, and serving optimizations; this breaks workflows when switching providers.
  • Some think Anthropic is slightly ahead on coding benchmarks; others say benchmark gaps are tiny and overshadowed by tooling and workflow design.
  • A few claim OpenAI models are now behind or “largely irrelevant,” while others still praise OpenAI’s image models and expect strong coding agents via Codex.

Organizational & Engineering Culture at AWS

  • Discussion of “SWAT teams” and Mantle is used to criticize Amazon’s broader culture: elite principal engineers can bypass bureaucracy, while lower levels are constrained by process and promotion‑driven empire building.
  • Others defend this as a pragmatic way to ship major new capabilities quickly, though acknowledging it does not scale to all work.

Ethics, Trust & Vendor Perception

  • Some organizations reportedly banned OpenAI but allowed Anthropic via AWS; this move may or may not change those legal/compliance stances.
  • Several commenters distrust OpenAI leadership and factor perceived ethics into tool selection; others argue competing founders are also intent on automating away human labor.

Impact on Market & Startups

  • Many see this as good for AWS‑centric teams (credits, unified billing, fewer vendors) and for Bedrock’s model diversity.
  • Some argue OpenAI is “scrambling” after missing revenue targets and may be 1–2 years late to the enterprise‑via‑cloud strategy.
  • There is concern that as enterprises standardize on cloud marketplaces like Bedrock, independent AI startups will find it harder to get in the door.